Fault Prediction Method of Boost Converter Based on Multi-Modal Components and Temporal Convolutional Networks

During long-term operation, power electronic converters are jointly affected by component degradation and operational disturbances, leading to pronounced nonstationary and multi-scale characteristics in output-voltage signals, which pose challenges for fault prediction. To address the degradation forecasting problem of Boost converter output voltage, this paper proposes a multi-scale temporal modeling method that integrates multivariate variational mode decomposition, distribution entropy-based complexity features, and a temporal convolutional network. Multivariate variational mode decomposition is employed to achieve frequency-aligned decomposition of the voltage signal, enabling effective separation of dynamic components at different scales. Distribution entropy is then introduced to characterize the evolution of local structural complexity in each mode, and multi-channel complexity feature sequences are constructed accordingly. Based on these features, a temporal convolutional network is used to perform unified modeling of short-term fluctuations and long-term degradation trends. Experimental results demonstrate that the proposed approach achieves consistently high accuracy across multiple independent runs, with average RMSE ranging from 0.0111 to 0.0179 and average MAPE from 1.15% to 1.84%. The low standard deviations further confirm its robustness for degradation trend prediction under varying operating conditions.